Findings

Findings is where you go from “here are a lot of traces” to “here are the specific behaviors worth looking at.” That can be a failure, an accuracy gap, or just a behavior you want to watch. Behind the simple search box, Findings is a multi-stage analysis system, not a single model pass. That depth keeps the noise down and the coverage high: real issues instead of false alarms, the subtle ones as well as the obvious, already grouped into clear patterns you can act on, and steady enough to trust over time. Across a wide range of issue types it reaches over 82% accuracy at surfacing real issues.

​

Three ways to work

There are three ways to work with your traces, and they build on each other:

• Discovery: investigate your traces at scale and surface patterns you didn’t know to look for.
• Tracking: pin the patterns that matter and watch how often they come back.
• Traces: browse the individual traces underneath and route them onward.

Discovery and Tracking are the two tabs of Findings; Traces is its own page alongside them. The usual path is to run a Discovery search, pin what matters into Tracking, and drop into individual Traces whenever you need the detail.

​

Traces

Traces shows ingested traces with the relevant data (model, latency, tokens, preview). Bulk-select any subset and route it onward in one of three directions:

• Assign review sends the traces to a Review Queue or to specific reviewers.
• Add to dataset → Evaluation Dataset turns them into ground truth for judges.
• Add to dataset → Generation Seed Dataset turns them into seed examples for synthetic generation. Pick whether each trace contributes its input only or its input and output, then create a new seed dataset or append to an existing one. The result is an ordinary seed dataset: create a spec from it and generate, exactly as you would from uploaded files.

The same actions are available on a single trace’s page and on a finding’s trace list, so you can route from wherever you found the behavior.

​

Diagnose a trace

The built-in Copilot helps you dig into a single trace, with an understanding of everything around it: the trace itself, any rubrics that apply, attachments, the findings it belongs to, past reviews, its user journey, and saved memory from earlier reviews. Ask it to summarize the trace, say what the model got wrong, or check whether the final answer matches the query. You can also point it at one part, like the query, the prompt, a single step, or the final answer.

​

Getting traces in

Findings connects to your observability tool (Langfuse or LangSmith today) and pulls real LLM traces in. See Connect AI Observability for setup, and Signals & Journeys if you also want user events correlated with those traces.

​

Next steps